On the acceleration of gradient methods: the triangle steepest descent method

Abstract: The gradient type of methods has been a competitive choice in solving large scale problems arising from various applications such as machine learning. However, there is still space to accelerate the gradient methods. To this end, in this paper, we pay attention to the cyclic steepest descent method (CSD), and prove that the CSD method has a gradient subsequence that is R-superlinearly convergent for the 2-dimensional strictly convex quadratic case. Moreover, we propose a new gradient method called triangle steepest descent method (TSD) which has a parameter $j$ to control the number of cycles. This method is motivated by utilizing a geometric property of the steepest descent method (SD) method to get around the zigzag behavior. We show that the TSD method is at least R-linearly convergent for strictly convex quadratic problems. The advantage of the TSD method is that it is not sensitive to the condition number of a strictly convex quadratic problem. For example, it performs better than other competitive gradient methods when the condition number reaches 1e20 or 1e100 for some strictly convex quadratic problems. Extensive numerical results verify the efficiency of the TSD method compared to other types of gradient methods.